↓Skip to main content
  1. Posts/

Self-Managing Knowledge Base: RAG System That Tracks Documents Automatically

·2 mins
Case Studies Rag N8n Supabase Knowledge-Base Claude-Code Coda
Table of Contents

I was drowning in scattered files. Google Drive, PCloud, Documents, Coda, Keep. My filing systems would degrade from organised to chaotic after abandoning various projects, creating significant retrieval challenges.

So I built BizBrains — a self-managing knowledge base where a structured table acts as the remote control for everything.

The Core Idea #

Most RAG implementations passively store vectorised content without tracking or removal. Add a document, it gets vectorised. Done. But what happens when you update a document? Or delete it? Most systems leave orphaned data in the vector store.

BizBrains V1 used a Google Sheet to track documents alongside their titles, descriptions, tags, and categories. When PDFs were added to a specific folder, an n8n workflow vectorised them and stored them in Supabase. Changing a document’s status to “Remove” automatically deleted its chunks from the vector store.

BizBrains V1 Architecture #

  • Document Tracking: Google Sheet managing vectorisation status
  • Ingestion: PDFs added to Google Drive folder, processed via n8n workflow
  • Metadata: Title, description, tags, category tracked per document
  • Removal: Status column automation for chunk deletion from Supabase
  • Retrieval: Cohere Reranker checking against metadata for accuracy

BizBrains V2 — Rebuilt with Claude Code #

In April 2026, I used Claude Code to recreate the system in one week, consolidating from six workflows to four. The Coda table now serves dual purposes: document source and workflow controller.

Coda BizBrain Document Controller

Additional V2 features:

  • Document Types: Knowledge documents are chunked and vectorised; Guides are stored whole with metadata-only vectorisation, returning links to original documents
  • Audience Segmentation: A column for future internal/external sharing control
  • User Interface: Coda buttons enabling status changes; Instant Webhook Chrome Extension triggering n8n workflow checks
  • Frontend: Query interface for searching the vector store

BizBrains V3 Frontend Design

Roadmap #

  • Internal Tags System: Using <int></int> tags to redact sensitive content in publicly shared documents — visible only to authenticated users
  • User Authentication: Required for delineating internal vs external access
  • Online Table Migration: Moving the Coda table into the frontend after authentication
  • Folder Tree View: Hierarchical metadata field for document organisation
  • External File Vectorisation: Supporting DOCX, Google Slides by vectorising URLs as metadata while files remain in their original locations